SAR target recognition via incremental nonnegative matrix factorization with Lp sparse constraint

In this paper a novel incremental nonnegative matrix factorization (INMF) for SAR target recognition is proposed in order to overcome the defect that conventional methods have in online processing. When training samples increase, unlike conventional NMF-based methods computing both original and new samples to retrain a new model, INMF just computes the new samples to update the trained model incrementally, which can avoid repetitive learning original samples and reduce the computation cost. Meanwhile, the INMF with L
_p
constraint is proposed by setting the updating process under L
_p
sparse constraint for matrices decomposition. The proposed L
_p
-INMF can solve the problem of the computational cost increasing along with the sample increasing which is common in traditional methods. Experiment results on MSTAR data verify that the recognition performance obtained by L
_p
-INMF outperforms other traditional methods, and the recognition efficiency can be improved.